Improving the unmanned aerial vehicle (UAV) image object detection algorithm based on YOLOv8s

There are several challenges that need to be overcome in UAV target detection tasks: firstly, there are a large number of small target objects in the images captured by the UAV; secondly, the size of the target objects may vary considerably; and lastly, the target objects may be occluded by other objects. To address these challenges, this research focuses on how to enhance the YOLOv8 algorithm by incorporating functional modules from the recently released YOLOv9 algorithm. These improvements aim to increase the accuracy and efficiency of UAV target detection under complex conditions. The improvements involve replacing the C2f module in YOLOv8 with the RepNCSPELAN4 module from YOLOv9, aiming to optimize computation, enrich gradients, and facilitate efficient feature aggregation. Additionally, a dedicated detection head targeting small objects is introduced, which further fuses shallow and deep-level feature information to greatly reduce the rate of missed detections and enhance detection accuracy for small targets. Furthermore, the global attention mechanism (GAM) is integrated into the YOLOv8 backbone network, leveraging the global information features acquired through UAV imaging. To overcome limitations in traditional loss functions for bounding box regression and to address the suboptimal optimization when handling predicted bounding boxes with the same aspect ratio but different width and height values compared to ground truth bounding boxes, the MPDIoU loss function is introduced. On the VisDrone2019 dataset, the improved version of the YOLOv8 model (YOLOv8-rgm) shows excellent experimental results. Compared to the original YOLOv8s algorithm, the average detection accuracy of the YOLOv8-rgm algorithm ([email protected], [email protected]) is improved by $9.2 \%$ and $6.6 \%$, respectively. This improvement algorithm shows superior performance compared to the newly released YOLOv9 algorithm and holds promising applications in UAV detection scenarios.